Proposal: Artificial Intelligence
I have a few concerns about the AI proposal.
1. The site content doesn't match the description.
The AI.SE proposal has the description
Beta Q&A site for conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment.
which does not really describe the kinds of questions that the AI.SE website receives. Instead of being about "life and challenges in a world where "cognitive" functions can be mimicked in a purely digital environment", the AI.SE website attracts many questions about how to implement intelligent agents, such as machine learning, neural networks, and reinforcement learning. A question about how to build a neural network to achieve a specific task, or how U-Net works, or how to express a neural network using linear algebra, is very different from a conceptual question about life in a world where cognitive functions are mimicked in a purely digital environment.
2. There is a large overlap with stats.SE
I appreciate that the popular imagination does not see a strong connection between "boring, old statistics" (p-value and t-tests and regressions) and "new, sexy, cutting-edge" artificial intelligence, but the fact is neural networks are a topic in statistics, broadly conceived. The stats.SE website already has deep coverage of machine learning and neural networks. Machine learning is even listed as a topic that users can ask questions about in the Stats.SE help page.
A number of the AI.SE topics on these questions are perfectly on-topic on stats.SE and are already answered.
The large overlap with stats.SE seems to cannibalize and existing website, which is undesirable. The SE blog has written about this in the past.:
Area 51 was always envisioned as a tool for broadening our scope — for creating new sites serving new topics and answering questions that were previously considered off-topic on our existing sites. Area 51 was never intended as a tool for creating overlapping sub-sites that would cannibalize users from our existing sites!
The problem of site overlap could have been avoided. A hypothetical Area 51 proposal for a website about the philosophical and conceptual components of AI could have been written. If this proposal excludes statistical topics in AI, this website would not overlap with stats.SE. Or, alternatively, the existing AI.SE website could conform to the description above. But that's not what's happened, and now AI.SE has a lot of overlap with stats.SE.
As moderator and experienced user of stats.SE, and as a professional who focuses on machine learning and neural networks, I can confidently say that a large portion of the questions on the AI.SE website are on-topic on stats.SE.
Here are some examples of overlap. I hope this makes it clear that the AI.SE site and stats.SE site have a lot of material in common. Both sites have thousands of questions, and it takes time to correlate content between them.
- AI.SE has a question What is the use of convolutional layer in expansion path of U-Net?
This question is on-topic on stats.SE because it asks how a specific neural network works. We have 85 questions about U-Nets on stats.SE. https://stats.stackexchange.com/search?q=u-net
- AI.SE has the question Correct way to work with both categorical and continuous features together This question is already answered on stats.SE, in several forms:
- https://stats.stackexchange.com/questions/479433/combining-categorical-and-continuous-features-for-neural-networks
- https://stats.stackexchange.com/questions/306610/neural-network-mlp-for-regression-with-3-continuous-features-1-categorical
- https://stats.stackexchange.com/questions/139129/how-to-recode-categorical-variable-into-numerical-variable-when-using-svm-or-neu
- AI.SE has the question How to select number of hidden layers and number of memory cells in an LSTM?
- What is the purpose of an activation function in neural networks? Stats.SE has several variations on this.
- https://stats.stackexchange.com/questions/228296/what-is-the-purpose-of-a-neural-network-activation-function
- https://stats.stackexchange.com/questions/333700/to-what-exactly-does-the-term-activations-refer-in-neural-networks
3. The overlap is contentious and confusing.
And the issue of overlap between AI.SE and Stats.SE has arisen several times on stats.SE. Speaking for myself, I am somewhat perplex that another website has been created which, in effect, overlaps with much of the stats.SE content.
- https://stats.meta.stackexchange.com/questions/5514/overlap-with-ai-stackexchange-and-datascience-stackexchange
- https://stats.meta.stackexchange.com/questions/5573/questions-on-reinforcement-learning-topics-would-be-better-suited-for-ai-stack-e?noredirect=1&lq=1
- https://stats.meta.stackexchange.com/questions/6050/should-stats-se-encourage-rl-questions-to-be-asked-on-ai-se
- https://stats.meta.stackexchange.com/questions/4243/after-data-science-se-artificial-intelligence-se
- https://stats.meta.stackexchange.com/questions/2094/where-should-i-ask-a-question-about-artificial-intelligence
- https://stats.meta.stackexchange.com/questions/5913/can-we-ask-questions-related-to-deep-learning-ai/5914#5914
4. Should machine learning questions on AI.SE be part of an existing site, or its own site?
The Area 51 FAQ provides a general rubric for answering this. I've substituted in X=AI.SE and Y=Stats.SE to make the general formula in the FAQ easier to read, and revised the focus of the general formula to the particulars of the topics in question. I know that, in this edited form, this is not the Official Area 51 rubric; however, I think it is illustrative of the core problem, which is that the scope of AI.SE is so expansive that it overlaps with existing websites.
In general, if a site makes sense as part of a bigger site, it's better to have one big site than a bunch of little niche sites. Machine learning questions on AI.SE should be subsumed by Stats.SE if:
- Almost all machine learning questions on AI.SE questions are on-topic for site Stats.SE
This criterion is established in header 2 on this post.
- If Stats.SE already exists, it already has a tag for machine learning, and nobody is complaining
This criterion is established by the meta threads on stats.SE about machine learning as a topic make it clear that stats.SE welcomes statistical questions about machine learning. This is header 3 of this post.
- You're not creating such a big group that you don't have enough experts to answer all possible questions
This one is hard to judge. The statistics reported on the AI.SE Area 51 page are mostly good, but the answer rate and number of daily questions need some work. This answer from DuttaA on Meta.SE describes a need on AI.SE for more experts:
Our site is very small and non-rewarding, there have been multiple experts who have came and gone (but hangs out in other SE sites). So this has resulted in a severe lack of community moderation.
Returning to the FAQ's criteria,
- There's a high probability that users of site Stats.SE would enjoy seeing the occasional question about machine learning.
Stats.SE has tens of thousands of posts about machine learning and neural networks. These questions are not occasional. Stats.SE has occasional posts about the history of these topics.
Stast.SE does not have many questions at all about philosophy of intelligence, theories about general intelligence in digital systems, or "conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment." Such questions are likely off-topic on Stats.SE but they may be on-topic on a Stack Exchange site other than AI.SE.
The AI.SE website, as it exists now, does not seem to meet the criteria required to stand on its own.
What to do?
I don't know what can be done, at this juncture, since the AI.SE website is about 4 years old. In all honesty, I'm not familiar with how Area 51 works, in specific terms.
It's clear that the site has evolved over that time, since its content and description differ. But I am concerned that right now, AI.SE and Stats.SE are a Venn diagram with an enormous overlap, and this seems to harm both websites.
The overlap isn't necessary. The AI.SE description is very different from the AI.SE content that exists today. An AI.SE website which grapples with "conceptual questions about life and challenges in a world where cognitive functions can be mimicked in a purely digital environment" is a novel and intriguing SE website proposal. I think it could be a nice community.
Here's another way to frame the issue. Suppose some new, unnamed site started with a very clear, non-overlapping description that AI.SE used. Then this unnamed site extended that description to also include the kinds of questions about theories of intelligent systems, pathfinding problems/algorithms, tree search or graph search algorithms and logic-based AI approaches. That would seem to be fine because there's still no overlap with stats.SE, and these questions all have a coherent home.
But what has happened instead is that the AI.SE proposal defined itself in terms that clearly have no overlap, and on the basis of this description, this website was able to get launched via the Area 51 process. Then, the AI.SE website changed its topics to have large overlap with existing websites, such as stats.SE. The effect of this metamorphosis is that there's a lot of confusion and duplication of effort, as well as a fracturing of expertise across different websites.